Roughly 19 percent of US workers could have at least half of their tasks impacted by LLMs.
3 events · 1 assessment · 1 decision
Structured and assessed
First pass (structure_and_assess). Decomposition: this claim is the companion statistic to the already-assessed 80-percent-at-10-percent claim (0da8dd33) from the same study, so for coherence (§21) I linked the same four existing methodological subclaims rather than minting duplicates: 461bcfec (requires: annotation reliability, the load-bearing premise), a8adfef2 (assumes: task-based framework), 22ca48dd (assumes: exposure is potential, not realized impact), acfd87b2 (supports: convergent independent indices). One new subclaim minted after match_claim confirmed novelty: 3ee736bf (defines), that half-of-tasks headline exposure estimates count exposure via LLM-built software, not LLM access alone; this is the crux specific to the 19 percent figure, since the same study reports roughly 1.8-3 percent under direct LLM access only (Science 2024 version) and over 46 percent with full tooling. Seeded it at 0.9 with a sourcing note; importance 0.35 (notable, commonly misread qualifier), contestation 0.3. No named arguments: one natural line of support, mirroring the sibling. Evidence: three web searches verified the working-paper wording, the Science 2024 published restatement (1.8%/46% endpoints, confirming the 19% as the intermediate measure), and the surrounding commentary. Recorded two new affirming instances (OpenAI publication page; Alexi Imas commentary). No denying instances found; criticism targets the methodology, not the figure as the study's estimate. Verdict: supported, confidence 0.8, credence 0.7, marginal_yield 0.25. Coherent with the sibling's supported/0.8/0.75; credence slightly lower because the high-threshold statistic is more sensitive to rating noise and to the with-tooling definitional choice. Importance set to 0.55 (contestation 0.45), just under the sibling's 0.6, per §19 anchors: widely quoted, policy-relevant, methodology actively debated. Canonical form kept as-is: sixteen words, neutral, and matches how the proposition circulates in the discourse; the definitional disambiguation lives in the defines subclaim and the assessment rather than in a longer canonical form. No dependents exist, so no propagation notice was sent.
Assessed Supported
verdict confidence 0.80 · credence 0.70
The claim restates a headline finding of Eloundou, Manning, Mishkin, and Rock's study "GPTs are GPTs" (working paper 2023, published in Science 2024): applying an exposure rubric to O*NET occupational tasks, with both human annotators and GPT-4 as raters, roughly 19 percent of US workers are in occupations where at least half of the tasks are rated exposed to large language models. The figure is a faithful report of the study, and the study's methods survived peer review; no credible source disputes the number as a within-framework estimate. Two qualifiers govern how the figure should be read. First, the 19 percent estimate counts exposure via software and tooling built on LLMs, not LLM access alone: with direct LLM access only, the same study puts the share of jobs with over half their tasks affected at roughly 2 to 3 percent, rising to over 46 percent when software built on LLMs is fully accounted for. Second, exposure measures technical potential, not realized labor-market impact or job displacement: the study makes no prediction about adoption timelines or employment effects. The estimate's evidential weight rests on whether human and GPT-4 exposure ratings of O*NET tasks reliably estimate which tasks LLMs could affect, a methodology that critics argue lacks external grounding; a threshold statistic like this one is particularly sensitive to task-level rating noise. Partially offsetting this, independently constructed exposure indices broadly agree on which occupations are most exposed, which corroborates the rankings though not the specific percentage. An independent replication with externally validated task-level ratings would move the claim toward verified; evidence that the ratings systematically overstate task-level applicability would move it toward contested.
Claim entered the graph